A short-term photovoltaic power interval forecasting method based on fuzzy granular computing and CNN-BiGRU

Yan Shi,Luxi Zhang,Siteng Wang,Wenjie Li,Renjie Tong
DOI: https://doi.org/10.1093/ijlct/ctad131
IF: 2.3
2024-01-01
International Journal of Low-Carbon Technologies
Abstract:Abstract This paper presents a short-term PV power interval prediction method combining fuzzy information granulation and CNN-BiGRU model. First, historical data of PV power generation is processed using fuzzy information granulation to determine the interval range. Subsequently, a CNN-BiGRU model is constructed, where the CNN module extracts local features of the interval range and the BiGRU module captures temporal patterns. The interval range is then fed into the CNN-BiGRU model for training, which enables accurate prediction of short-term PV power generation intervals. Finally, the predicted power generation interval ranges are given. The experimental results show that this hybrid interval prediction method is first applied to the prediction of short-term PV power scenarios, which fully utilizes the advantages of information granularity to express the data intervals and CNN-BiGRU to learn the complex time series patterns, and realizes the accurate and reliable prediction of short-term PV power intervals. Compared with direct numerical prediction, the method provides an expression of prediction uncertainty, which provides an important reference for the operation decision of PV power plants.
energy & fuels,thermodynamics
What problem does this paper attempt to address?
### Problems the paper attempts to solve This paper aims to address the challenges of short - term interval prediction in photovoltaic (PV) power generation. Specifically, the research focuses on the following aspects: 1. **Uncertainty and volatility of PV output**: The significant intermittent and volatile characteristics of PV systems bring uncertainty to power grid operation, which poses a challenge to the stability of the power grid. Therefore, improving the accuracy and controllability of PV power prediction is of great significance for optimizing the energy structure and achieving the development goals of being eco - friendly and carbon - neutral. 2. **Limitations of existing prediction methods**: Most of the current research on PV power prediction mainly focuses on accurate numerical indicators rather than providing prediction intervals. Although numerical prediction can give specific predicted values, it cannot express the uncertainty of prediction. Interval prediction, on the other hand, can provide the distribution interval of prediction errors, thus better supporting decision - making. 3. **Modeling of complex time - series features**: The existing interval generation methods are relatively simple and fail to fully utilize the advantages of deep neural networks in capturing the temporal distribution of time - series data. Therefore, a method that can effectively model complex time - series features is required. ### Solutions To address the above problems, this paper proposes a short - term PV power interval prediction method based on fuzzy information granulation and the CNN - BiGRU model. The specific innovations are as follows: 1. **Fuzzy information granulation**: Use fuzzy information granulation technology to express the interval nature of PV power. Construct information granules through fuzzy set theory and convert the original numerical sequence into a fuzzy granular form to meet the needs of interval prediction. 2. **CNN - BiGRU model design**: Design an end - to - end feature learning model that combines a convolutional neural network (CNN) and a bidirectional gated recurrent unit (BiGRU). The CNN module extracts local features, and the BiGRU module captures time patterns, achieving in - depth feature extraction of historical data and effective modeling of complex non - linear data. 3. **Direct prediction of distribution intervals**: Directly predict the distribution interval of PV power instead of a single numerical prediction. This helps to capture the inherent uncertainty of prediction and provides valuable references for operational decisions. ### Experimental results The experimental results show that the proposed FIG - CNN - BiGRU model outperforms other models, such as FIG - BiGRU, FIG - LSTM, FIG - SVM, and FIG - XGBoost, in performance indicators such as probability of coverage (PICP), prediction interval normalized average width (PINAW), and prediction interval normalized average deviation (PINAD). Specifically: - **Probability of coverage (PICP)**: For example, in the January data set, the PICP of the FIG - CNN - BiGRU model is 3.6% higher than that of FIG - BiGRU, 9.0% higher than that of FIG - LSTM, 5.0% higher than that of FIG - SVM, and 4.1% higher than that of FIG - XGBoost. - **Prediction interval normalized average width (PINAW)**: Also in the January data set, the PINAW of the FIG - CNN - BiGRU model is 1.5% lower than that of FIG - BiGRU, 2.7% lower than that of FIG - LSTM, 2.1% lower than that of FIG - SVM, and 4.4% lower than that of FIG - XGBoost. These results indicate that the proposed model can more effectively capture the fluctuation characteristics of time series, thereby improving the prediction accuracy of PV power intervals.